Browse Chinese AI Skills
Search and filter 212 entries by workflow. Thirteen are maintained in this repository; third-party entries are catalog links, not endorsements.
Open the searchable directory →CHINESE AI SKILLS · CODEX · CLAUDE CODE · CURSOR
Chinese Agent Skills you can inspect, install, and reproduce.
A maintained directory for Chinese-language workflows—not just a link dump. It separates catalog entries, first-party Skills, installation evidence, automatic activation, and final task outcomes.
Evidence boundary: Claude Code and Cursor task-level compatibility are still untested by this repository. Installation, discovery, activation, and task completion are separate claims.
01 / START
Four useful paths, depending on what you need
Try one first-party Skill
Choose one of 13 natural tasks, install only that Skill, restart the client, and compare your result with recorded evidence or preregistered gates.
Choose a first task →Review before install
Use the read-only local scanner to flag scripts, symlinks, network references, credential terms, file mutation, and high-attention commands.
Read the scanner limits →02 / INSTALL
Install one Skill before installing everything
Inspect which Skills the repository exposes.
Read SKILL.md and any scripts or references.
Limit the first test to the workflow you need.
Use a natural task without naming the Skill.
npx skills add https://github.com/sanhuang520-ship-it/awesome-chinese-ai-tools --skill chinese-typography -gThe repository recorded all 13 first-party Skills being selected in one Codex environment. Ten completed the recorded task, one stopped for required input, and two large tasks failed before narrower retests passed. Two later preregistered failures were fixed and replayed with the unchanged original task. These observations are not an overall accuracy rate or a cross-client guarantee.
03 / VERIFY
Evidence is useful only when the failure stays visible
Compatibility method
Separate CLI discovery, file installation, automatic activation, task completion, and environment blocks.
Read the four-layer method →Failure repair guide
Do not rewrite the task, lower the acceptance gate, or overwrite the first failure. Make the smallest instruction fix and replay the same task.
See two 3/4 → 4/4 cases →Reproducible cases
Seven historical tasks preserve their exact wording; six summary-only records are not reconstructed after the fact.
Copy a verbatim task →Contribute one real result.
Success, failure, non-activation, and environment blocks are all useful when version, task, outcome, and limits are recorded.